Company Worlds

A copy of your companyfor testing AI agents.

By invitation, for Butterfly clients

One part of your business, rebuilt in software: its systems, its teams and its working week. AI agents would do the job there a hundred times over before they meet a customer.

Kinds of work in every companyPick one to bring it round and see its tasks

01 · How it works

What a Company World is,and how it works.

A working replica of one area of your company, such as motor claims or month-end close. It would run in your cloud, in software, so AI agents could do the job there again and again before they do it for real.

01Copy one areaWorking copies of its 3⁠–⁠6 systems, simulated staff in 5⁠–⁠15 roles, its working week and the KPIs you already report.What is inside
02Prove the copyReplay last quarter’s real inputs. The World has to reproduce last quarter’s KPIs, within a tolerance you agree, before it scores any agent.The Fidelity Score
03Run it a hundred timesEach candidate agent works the same hundred seeded weeks, with their surges, outages and fraud.Watch one week
04Score every taskRight every time, not right once; a P&L against your own team; every failure logged.The verdicts
05Approve, then retestThe task owner approves each task’s blueprint. It would run again every week, and fully each quarter.A blueprint

02 · What is inside

Not a crystal ball.A working replica.

Three ideas get mixed up. A Company World takes its loop from the first and its shape from the second, and is never the third.

01
An environment
Software with a state, the actions an agent may take and a score. Reset it, run it again. The training ground AI labs use.
02
A digital twin of an organisation
A living software model of how a business works: its teams, its systems, the way work flows between them.
03 · Not this
A learned simulator
A neural network that generates scenes, such as the world models made for robots and games. Useful there; no way to test a claims desk.

A Company World is an environment shaped like a digital twin: rule twins of 3–6 systems, a simulated clock, an event queue, simulated staff in 5–15 roles and the KPIs you already report.

Rule twinsof 3–6 systems
A simulated clockMon 09:00 – Fri 17:30
An event queuearrivals, surges, outages
Simulated staff5–15 roles
KPIsthe ones you already report

A World Week holds hundreds of tasks. Each agent works them one at a time, as your team would, and every one is scored.

20.6%

The top agent completes about one in five real computer workflows of around 1.6 hours.

Source: OSWorld 2.0, arXiv, 28 Jun 2026 ¹²

$9,235 vs $11,182

A frontier model that leads the Remote Labor Index earns less than its predecessor running a simulated vending business.

Source: Andon Labs, 24 Sep 2026 ¹⁷

2.5% → 21.25%

The top agent’s share of real freelance projects, a year apart. The ground moves every quarter, so the test has to move with it.

Source: Scale AI, Remote Labor Index, fetched 25 Oct 2026 ¹³

So test in yours.

03 · The Fidelity Score

No agent is scoreduntil the World passes.

Replay last quarter’s real inputs. The World has to reproduce last quarter’s real KPIs within a tolerance you agree, or it is fixed and replayed until it does.

Illustrative

Cycle timedays

Real 9.4World 9.2Not replayedWithinOutside

Claims leakage£k a week

Real 14.2World 14.9Not replayedWithinOutside

Reopen rate%

Real 6.1World 8.4Not replayedWithinOutside

Complaints per 1,000 claims

Real 9World 9.5Not replayedWithinOutside

0 of 4KPIs reproduced. Press 01 to replay last quarter.

Last quarter, real
The World’s replay
Agreed tolerance

04 · One week, a hundred times

One week of motor claims,run a hundred times.

A vendor’s agent works a simulated week: Monday’s backlog, Tuesday’s outage, Wednesday’s storm, Thursday’s fraud ring. Scrub through it, then replay any of the hundred weeks.

World Week 1 of 100

Your company · motor claims · agent under test: a vendor’s agent

Example

The week opens with 50 claims waiting from the weekend.

Claim waiting
Being worked
Storm claim
Linked claim (fraud ring)
Sent back to a person
Wrong, unnoticed: the World sees it

A hundred weeks

One line per World Week, one row per task. 8 meltdowns flagged. Pick a week to replay it.

FNOL triageClears the bar · 95 of 100

No more than 2 misroutes a week, and faster than the team’s 38 minutes.

Fraud referralWith a person checking · 70 of 100

It joined the fraud ring’s dots. Its referrals are worth having; an investigator checks every one.

Coverage decisionsNot yet · 17 of 100

Matched the team’s 97%, and melted down in the outage or the storm in 8 of 100 weeks. It goes to Butterfly Lab for experiments, then a re-test.

Week 1: 96.8% of cover decisions right, fraud ring referred, £22k leaked

Passed the week
Failed the week
Meltdown

Monday 09:00. 0 claims open, 0 settled.

05 · The verdict

Clears the bar, with a person checking,or not yet.

A World would end in a decision for each task, taken from a hundred weeks, not one good afternoon. The agent that clears the bar gets a blueprint; the task owner approves it.

Clears the bar
FNOL triage
Passed 95 of 100 weeks: no more than 2 misroutes a week, and faster than the team’s 38 minutes.
With a person checking
Fraud referral
It joined the fraud ring’s dots in 70 of 100 weeks. Its referrals are worth having; an investigator checks every one.
Not yet
Coverage decisions
Matched the team’s 97% in 17 of 100 weeks, and melted down in the outage or the storm in 8. It goes to Butterfly Lab for experiments, then a re-test.

Clears the bar

Blueprint · FNOL triage

Motor claims, ground floor

Candidate: a vendor’s agent · 95 of 100 World Weeks passed

Reads: New notifications in the claims platform · call transcripts · the claims inbox · the policy’s cover summary
Writes: The new claim record · its route · a fraud flag
Never: Payments · coverage decisions · letters to customers
A person in the loop: The team leader reviews every fraud flag and every vulnerable-customer flag
Must keep passing: Retested every week, and fully each quarter, on the same hundred weeks and new ones
Approved by: The FNOL team leader: the task owner

Example

06 · Why your own

The model that wins one worldloses another.

Agents take on harder work every quarter. Few companies will hand one their sensitive processes on a leaderboard’s word: the test has to be their own.

$9,235 vs $11,182

A frontier model that leads the Remote Labor Index earns less than its predecessor running a simulated vending business.

Source: Andon Labs, 24 Sep 2026 ¹⁷

2.5% → 21.25%

The top agent’s share of real freelance projects, a year apart. The ground moves every quarter, so the test has to move with it.

Source: Scale AI, Remote Labor Index, fetched 25 Oct 2026 ¹³

So test in yours.

0706 · Example Worlds

Every areais a world.

Five Worlds, each drawn up from its area’s own systems, roles, tasks and KPIs. Motor claims and month-end close come first: claims are the heart of an insurer, and every company closes its books.

Motor claims department

Illustrative design

One of the two natural first Worlds.

Systems

Claims platform
Policy admin
Payments
Repairer portal
Telephony
Email

Roles

FNOL handler
Claims handler
Engineer
Fraud investigator
Team leader
Simulated customers and repairers

Example tasks

FNOL triage
Coverage check against policy wording
Repair authorisation
Total-loss settlement
Fraud referral

KPIs

Cycle time (days)
Claims leakage (£)
Reopen rate (%)
Complaints per 1,000 claims

0807 · What a World holds

From £250k:one area, made to measure.

1area
3–6systems
5–15roles
100–300tasks, each approved by its task owner
~12weeks to make

One illustrative structure

Draft for review · subject to contract

£200k: The replica, and the first 100 World Weeks for each candidate agent
£50k: Payable when the Fidelity Score passes
Included: Inference: Butterfly would pay the model bills, and show you what they came to

For Butterfly clients, after a Readiness Diagnostic (£60k). It would run in your cloud; Butterfly staff would have zero access.

What you receive

01The £-ranked map of the area: measured frequency × estimated £ cost for each task
02The World itself, running in your cloud
03At least 100 seeded World Weeks per candidate agent
04A P&L for each agent against the human baseline
05For each task, how often it is right every time, not right once
06A failure taxonomy and a log of safety incidents
07A blueprint for each task
08Retested every week

Candidates side by side: the vendor’s agent, your own agent, and frontier models at different effort levels. AI judges rank them; the people who do the job confirm the £ figures.

0908 · How to start

Worlds are by invitation,for Butterfly clients.

Every World starts with the Readiness Diagnostic: your top tasks, priced, and the first ones tested.

Now
Company Worlds are made by invitation, for Butterfly clients.
Next
The first World, made with a Butterfly client: motor claims or month-end close.
Then
More areas and more clients, once the first World has passed its fidelity backtest.

Your proposal

Get your proposal.Then register your interest.

A readiness report made from your company’s public website: what your tasks cost, and the first 25 to test. If it reads right, register your interest: a person from Butterfly will call you to talk it through.

Or read on:Agent benchmarks How it works